Deep Learning Kinematics

Differentiable Forwad Kinematics for TensorFlow and Keras

Supported Joint Types:

• Fixed
• Revolute
• Continuous
• Prismatic
• Floating - 6-DoF joint (not coverd by unit tests)
• Planar (not coverd by unit tests)

Installation

Install from PyPi

$ pip install dlkinematics

Usage:

import tensorflow as tf
from dlkinematics.urdf import chain_from_urdf_file
from dlkinematics.dlkinematics import DLKinematics

# Load URDF
chain = chain_from_urdf_file('data/human.urdf')

# Create DLKinematics
dlkinematics = DLKinematics(
   chain,
   base_link="human_base",
   end_link="human_spine_2",
   batch_size=2)

# Joint configuartion
thetas = tf.Variable([1., 2., 3., 4.], dtype=tf.float32)

# Forward pass
with tf.GradientTape() as tape:
    result = dlkinematics.forward(thetas)

print(result)
print(tape.gradient(result, thetas))

As Keras Layer

from dlkinematics.training_utils import ForwardKinematics
from tensorflow import keras
import tensorflow as tf

model = keras.Sequential()

FK_layer = ForwardKinematics(
   urdf_file = 'path/to/urdf',
   base_link = 'link0',
   end_link = 'linkN',
   batch_size = 2)

model.add(FK_layer)
# Output shape of FK_layer is (batch_size, 4, 4)

Run tests

The tests use ROS packages to validate the result of the dlkinematics module.

1. Build the docker image for tests:
   $ docker build -t dlkinematics_tests .

2. Start the container in the root folder of the project:
   $ docker run -it -v $PWD:/work dlkinematics_tests python3 -m pytest

3. Execute all tests:
   $ docker run -it -v $PWD:/work dlkinematics_tests python3 -m pytest
   Execute only a single testfile:
   $ docker run -it -v $PWD:/work dlkinematics_tests python3 -m pytest tests/test_prismatic.py